Implementing circuit-cutting in CudaQ

[vineetbansal]

Hello - I'm tasked with implementing a quantum circuit-cutting algorithm in cudaq (decomposing a quantum circuit into subcircuits and using classical computation to "combine" their results). We have a reference implementation in qiskit but we'd like to move to cudaq to leverage multiple gpus.

My question is - what might be a possible roadmap for us to do this? qiskit provides a DAG view of the circuit object that one can manipulate (and thus create sub-circuits). I didn't find any such possibility in the CudaQ Python API. Is this correct, or maybe I'm missing something? We were hoping to do this in Python, but if dropping down to the C++/IR level is the only way to do this, we'd like to know..

The graph examples I see in cuda-academic are of course for solving classical graph problems using quantum circuits, so not what we're looking for.

Thanks!

[poojarao8]

Hi @vineetbansal

To clarify, are you primarily interested in wire-cutting or gate-cutting? Also, is your goal to identify the optimal cut locations automatically?

I don't have a low-level roadmap for circuit cutting within CUDA-Q, but I have an application-level gate-cutting (single-cut) implementation that leverages multiple GPUs via cudaq.observe_async and our mqpu backend. This approach could serve as a useful starting point for your work.

1. It starts with a predefined list of 2-qubit unitaries, with the cut location explicitly specified, rather than beginning with a DAG object.

2. At the designated cut location, I employ matrix decomposition functions to express the entangling matrix as a linear combination of Pauli terms.

3. Subsequently, two CUDA-Q kernels are constructed using the kernel builder, each representing one of the two partitions of the original circuit. These subcircuits take cudaq.pauli_word objects as input parameters (with ongoing discussions to enhance this support), and the remaining gates are added.

4. From each subcircuit, corresponding circuits are generated for the terms identified in the Pauli decomposition by passing these terms.

5. The expectation value for each of these resultant circuits is then computed in parallel using cudaq.observe_async.

6. Finally, the coefficients obtained from the Pauli decomposition are utilized to collate the results and reconstruct the overall expectation value.

Let me know if you have any further questions.